Going deeper – deep learning for heterogeneous treatment effects
Since modern deep learning started gaining traction in the early 2010s, we have seen a continuous progression of breakthroughs. From AlexNet (Krizhevsky et al., 2012), which revolutionized computer vision, through Word2vec (Mikolov et al., 2013), which changed the face of NLP forever, to Transformers (Vaswani et al., 2017) and modern generative architectures (e.g. Radford et al., 2021, and Rombach et al., 2022), which fueled the generative AI explosion of 2022-2023.
Although the core idea behind (supervised) deep learning is associative in its nature and, as such, belongs to rung one of the Ladder of Causation, the flexibility of the framework can be leveraged to improve and extend existing causal inference methods.
In this section, we will introduce deep learning architectures to model heterogeneous treatment effects (aka conditional treatment effects (CATE)). We’ll discuss the advantages of using...
